Nina Schick: Why 2025 Was the Last Normal Year Before AI Takeoff

Nina Schick argues that 2025 was the last normal year and 2026 is the year AI takes off, driven by faster capability, a US and China race to control it, and constraints that are as much about people as about power. In The Last Normal Year, a 44-minute talk to camera, the author and adviser lays out a framework called industrial intelligence: AI as a manufactured product with a physical supply chain, an energy bill, and a public that has not agreed to pay it.

TLDW

Schick tracks three signals. Capability keeps accelerating, with Anthropic’s unreleased Mythos model as the evidence. Competition between the United States and China is becoming the defining contest of the century, and only those two can build every layer from raw materials to applications. Constraints will decide both, and while chips, memory and electricity can be solved with money and time, the shortage of skilled trades and the collapse of public trust in Western democracies may not be. The closing argument is that the largest risk is not a rogue superintelligence. It is democracies deciding not to build.

Thoughts

Schick’s argument turns on a constraint that has nothing to do with chips: people. Most talk about AI bottlenecks stops at energy and semiconductors, and Schick grants that those will yield to capital. Then comes a number that is easy to picture. If roughly 90 gigawatts of data centers come online in the United States within two years, they need about 150,000 electricians, out of a national pool Schick puts at around 800,000, many of whom are not qualified for this work. The technology that is supposed to automate white-collar jobs is short of blue-collar ones. The second half of the people problem is consent. In a democracy, a public that does not want something can stop it, and Schick cites more than $100 billion of AI infrastructure blocked or delayed last year. Whatever you think of the politics, that is a sharper account of what could slow AI than another chart of GPU supply.

The talk is at its most honest near the end, where it runs into its own tension. Schick wants the public to see AI as “by the people, for the people,” and in the same passage concedes that frontier intelligence is expensive, scarce, and reaching mainly the companies that can pay Anthropic a million dollars a year. The proposed fix is to drive the price down until intelligence is distributed like a utility. That may happen. But it means the benefits arrive later than the costs, and the people being asked for patience are the ones watching data centers go up near their homes. The ratepayer pledge Schick praises, in which hyperscalers agree to bring their own power so household bills do not rise, is the one concrete example in the talk of solving that problem with a commitment and not a message.

The economics section is the most useful part for anyone running a business. The usual line is that AI is getting cheaper, and Schick agrees for old models: a million tokens through GPT-3 cost about $60 and the equivalent cost six cents by 2024. But the frontier has not followed. Models that reason for longer and act on their own burn far more tokens per task, so the best results stay expensive, and Schick describes developers spending $10,000 a day on agent experiments. “Cheap intelligence is cheap and valuable intelligence is extraordinarily expensive still.” The practical consequence is that a company’s output starts to depend on how many tokens it can afford, not only on how many people it employs. If Schick is right that supply will stay short, prices at the top could rise before they fall.

The five-layer pyramid is a good corrective to model-watching. Energy, raw materials and manufacturing sit at the bottom, then chips and compute, then models, then applications, then full integration into the economy. Schick’s line is that you cannot lead from the apex if you are hollow at the base, and the case that China is strongest exactly where the West is weakest is well made. Some of the figures deserve care. Schick flags the claim of 200 million skilled workers as one to take with a pinch of salt. The statement that half the world’s elite AI researchers are in China is close to a widely quoted figure about where top researchers were educated, which is not the same as where they work now. The direction of the argument survives those caveats. The point about open source is the one I had not considered: if American frontier models stay closed and Chinese ones are open, much of the world may end up building on the Chinese stack by default.

Where I part ways is the diagnosis of public opinion. Schick treats low trust mostly as a failure of persuasion: the builders have told the story badly, and talk of doom and of a “permanent underclass” has not helped. That is true as far as it goes. But people may be skeptical because they have listened carefully. They have been told, by the industry itself, that their jobs are at risk and that the gains will be concentrated. A better narrative will not fix that. Visible benefits will, such as cheaper care, better services, and work that pays. The closing warning, that the biggest risk is democracies failing to rise to the moment, lands harder if you read it as a demand on the builders and not on the public.

Key Takeaways

Schick’s three signals for separating signal from noise in AI: capability, competition and constraints.
“Industrial intelligence” means AI is an industrial process (sand to silicon to gigawatt-scale “intelligence factories”) and will eventually be distributed like a utility.
The predicted wall never came. When training data ran short, researchers scaled inference time, letting models think longer, and then autonomy, letting them act for longer.
Today’s best models were trained on older infrastructure. The first one and two gigawatt clusters, newer hardware and newer efficiency gains have not yet shown up in public models.
Schick presents Anthropic’s Mythos as a step change: withheld from public release, shared with partners under Project Glasswing, and able to find and exploit long-unpatched vulnerabilities in major operating systems and browsers.
The conclusion drawn from that is that cyber defense now requires AI. “You need AI to fight AI.”
Old-model prices fell 99.9%, but frontier intelligence remains costly because reasoning and agents use many more tokens. Cheaper tokens raised total demand for compute, which Schick calls Jevons paradox at scale.
Figures Schick cites: OpenAI projected to spend $25 billion on training compute this year and $121 billion by 2028, and Anthropic’s annualized revenue going from $1 billion at the end of 2024 to $9 billion a year later, with $120 billion projected for December 2026.
Of the 60 to 70% of businesses that say they deploy AI, Schick says fewer than 1% have mature deployments, and that over a thousand Anthropic customers spend more than $1 million a year.
Only the United States and China can be vertically integrated across all five layers. Taiwan (chips), the Netherlands (ASML’s EUV machines), South Korea (high bandwidth memory) and the Gulf (energy and compute) each matter at one layer.
On energy, Schick says China added more than 400 gigawatts of generating capacity in 2025, over 300 of it solar, and is building 38 nuclear reactors, while Germany shut its reactors down.
Hyperscalers are becoming power producers, with small modular reactors, restarted nuclear plants and on-site generation, because the grid cannot meet the demand.
Schick cites survey data putting public support for AI at about 87% in China and just over 30% in the United States, and predicts AI will be the top political issue by 2028.

Chapters

00:01 Three Signals: Capability, Competition and Constraints

Schick opens with the frame for the whole talk. Capability at the frontier is accelerating, competition between nations is reordering geopolitics, and constraints will decide both. The physical constraints are well known. The one not discussed enough, Schick says, is people: trust and talent. The process itself is described as modern alchemy, turning sand into intelligence.

03:05 Why AI Did Not Hit a Wall

Skeptics expected progress to stall once the internet’s text had been used up. Researchers instead spent compute on inference, so models could think longer and check their own logic, and reasoning improved. The newest axis is autonomy, with models acting independently for longer stretches. Schick is bullish on the curve while declining to call it AGI.

06:05 Mythos, Project Glasswing and AI to Fight AI

Mythos is Schick’s proof that the curve is steepening. Its cyber abilities were not trained for on purpose, and Anthropic chose to give it only to partners so they could find and patch weaknesses first. Because such abilities will spread to other models, Schick argues that secure systems without AI on the defending side are no longer viable.

09:11 The Economics of Frontier Intelligence

There are two price curves. Commodity intelligence has collapsed in price, and frontier intelligence has not, because it consumes so many tokens. Schick expects a supply crunch as more companies move past chatbots to real deployments, and suggests frontier pricing could rise in the short to medium term. The development of that frontier is, in effect, being subsidized by US hyperscalers, labs and venture capital.

15:18 The AI Superpower Race and the Five-Layer Pyramid

Schick starts from Putin’s 2017 remark that whoever leads in AI will rule the world, notes that China published its plan to lead by 2030 the same year, and that the United States made AI dominance explicit policy in 2025. The pyramid runs from energy and industry at the base to transformational integration at the top. Offshoring the industrial base is called one of the West’s great strategic blunders.

20:41 China’s Advantage at the Base

China is already the world’s factory, and Schick quotes Tim Cook’s point that Apple builds there for the talent and ecosystem, not the cost. A “China shock 2.0” is coming as the country leapfrogs in electric vehicles, batteries, drones and drug development. On models, China is close behind and leads in open source, helped by distillation, smuggled chips and a deep pool of researchers.

26:43 From the Petrodollar to the Compute Dollar

For fifty years oil set the terms of geopolitics. Schick argues intelligence, and so compute, now takes that place. Sovereignty rests on prosperity and security, both of which are downstream of AI. A state that cannot produce intelligence, or secure it through trusted supply chains, is giving up sovereignty.

28:15 The Physical Constraints: Chips, Memory and Power

Over 90% of advanced chips come from Taiwan and over 80% of high bandwidth memory from two South Korean firms, and some builds are waiting on transformers and switchgear. Schick says more than 400 gigawatts has been requested from the Texas grid, over 70% of it for data centers. Europe, a net energy importer, is at a disadvantage. Hyperscalers now treat power generation as part of their own stack.

34:20 The People Constraint: Electricians and Public Trust

Physical limits can be bought through. People cannot. Schick describes a shortage of electricians and plumbers and a deep trust deficit, citing a poll in which AI was less popular than ICE. The talk points to organized bipartisan resistance, a proposed federal moratorium on data center construction from Bernie Sanders and Alexandria Ocasio-Cortez, and reports of violence against people associated with the buildout.

40:29 The Intelligence Divide and the Risk of Not Rising

Schick asks what happens when the price of intelligence reaches zero, and admits it is nowhere near. The gap between those with frontier access and those without is growing, and feeding the backlash. China has a clear national mission, Schick says, and the West does not. The final warning is that democracies failing to act is a greater danger than superintelligence.

Notable Quotes

“We are manufacturing non-biological intelligence and those who are able to control the production and deployment of that intelligence are going to have an extreme and compounding edge both economically as well as from a security perspective.”
Nina Schick, on why the AI race is geopolitical

“Cheap intelligence is cheap and valuable intelligence is extraordinarily expensive still.”
Nina Schick, on the gap between commodity and frontier model pricing

“Tokens aren’t a perk, they’re a requirement for productivity and discovery.”
Nina Schick, on compute as a new input to output

“You cannot be an AI superpower. You cannot lead from the apex if you’re hollow at the base.”
Nina Schick, on the five-layer pyramid

“Just as the petro dollar set the terms for the geopolitical order for the last 50 years, the compute dollar will set the terms for the next 50.”
Nina Schick, on intelligence replacing oil as the central resource

“It might turn out that the biggest bottleneck to developing industrial intelligence isn’t energy. It isn’t semiconductors. It isn’t high bandwidth memory. It’s actually the electricians and the plumbers.”
Nina Schick, on the talent constraint

“The biggest risk is that democracies fail to rise to this moment.”
Nina Schick, closing the talk

The figures in this post are Schick’s, as stated in the talk, and several are projections. Watch the full talk here for the complete argument.